Skip to content

Research separate from news

What could become important next

A preprint is a signal, not a finished product or an independently confirmed result. We therefore track research papers separately and show the actual state of supporting evidence for each one.

1381 published research events

Latest work

A significant claim from a single source is published only after further confirmation.

Research only one source so far

Persistent memory only slightly improved the success rate of an agent for drawing Chinese characters

The study compared five variants of an agent on 668 tasks. Memory with free-form notes increased the first-attempt success rate by 4.6 percentage points, while memory supplemented with reusable code increased it by 2.8 points. The difference between the two designs was not statistically significant overall.

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Study examines the costs and limitations of evaluating AI agents

The authors of the study Emerging Trends in AI Agent Evaluation stress that evaluating agents requires checking the entire process. According to the authors, evaluation credits alone for several agents in the ExploitBench test cost tens of thousands of dollars.

AI Snake Oil (Narayanan & Kapoor, Princeton) — a critical counterpoint to hype Original source ↗
Research only one source so far

E(3)-Pose model uses symmetry to track fetal head movement during MRI

Researchers from MIT, Boston Children’s Hospital and Inria developed the E(3)-Pose model, which uses the symmetry of the fetal head to track its orientation during MRI. According to the authors, it is intended to help produce higher-quality 3D images; the available excerpt does not report numerical results.

MIT News – Artificial intelligence Original source ↗
Research only one source so far

Study introduces MetaCurate for selecting training data for coding agents

MetaCurate creates filters for training trajectories using pairwise LLM evaluations. According to the authors, it achieved the highest average SWE-Verified score among the methods compared in subsequent experiments; the search process for the comparison method involving repeated training was 3.5 times as costly.

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Study of MoE models shows which tokens most affect responses when the number of experts is limited

A study of four families of MoE models found that drops in F1 when limiting the number of experts are often caused by the response format. Keeping the full number of experts for about 4 % of tokens kept F1 within one point of the baseline in some comparisons. Time to first token fell by about 14 % for inputs with 11 thousand tokens.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Study: Routing between language models based on answer confidence depends on the type of task

A study of 42 288 predictions found that routing based on answer confidence helps with reading comprehension and topic classification, but does not deliver consistent improvements on knowledge and reasoning tasks. Differences between languages in forwarding requests to larger models reached 30.33 percentage points.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Study describes hallucination detection in Bengali using search and computation

The study combines search, evaluation by language models and Python execution to detect hallucinations in Bengali. Kaggle reports a macro-F1 of 0.956; however, an audit found overlap between the corpus and 38 evaluated inputs. The study does not measure the contribution of individual components under controlled conditions.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Study introduces BayesianPRM for evaluating LLM reasoning steps

The authors introduce BayesianPRM, which accounts for uncertainty in the evaluation model and the difference between training and deployment policies when evaluating reasoning steps. According to the authors, the method produced more accurate reward estimates and improved performance with test-time scaling in multimodal tests.

arXiv cs.LG (Machine Learning) Original source ↗